Cannot install on Apple Silicon

[msg-systems]

Cannot install on Apple Silicon

[KTRosenberg]

I'm not sure this is resolved. Coreferee installation on macOS using the ARM chips still fails. Are you planning on supporting that platform officially?
Also, do you anticipate supporting the latest Spacy?

[KTRosenberg]

I'm not part of coreferee unfortunately 😞 but I thought that Edit 4 suggested things were working now. I'm currently stuck trying to get it running myself.

No, unfortunately it's not resolved. I had to install from source and do a bunch of painful things externally using mac-specific "hacks." pip installation completely fails due to the tensorflow dependency, and I had to use conda. Also, the fact that only Spacy < 3.2 is supported seems to prevent gpu acceleration or something. I hope this can be resolved and potentially updated for compatibility with 3.2.

[philippnoah]

I think this can be closed now.

[philippnoah]

I'm not part of coreferee unfortunately 😞 but I thought that Edit 4 suggested things were working now. I'm currently stuck trying to get it running myself.

[philippnoah]

🎉 I just got it to work with:

miniconda==4.11.0
python==3.9
numpy==1.21.4
tensorflow-macos==2.7.0
spacy==3.1.0
coreferee==1.1.2

I had to install the other dependencies before coreferee, then run pip install coreferee==1.1.2 --no-deps. Hope this works for you too, let me know.

[KTRosenberg]

🎉 I just got it to work with:

miniconda==4.11.0
python==3.9
numpy==1.21.4
tensorflow-macos==2.7.0
spacy==3.2.1
coreferee==1.1.2

I had to install the other dependencies before coreferee, then run pip install coreferee==1.1.2 --no-deps. Hope this works for you too.

Could you show the full steps for installation of the environment and dependencies? I just don’t want room for ambiguity since I had to do a lot of crazy side maneuvers as shown above. I am also pretty sure I did something similar to what you did and the pip tensorflow package always failed.
Did you use miniforge for conda as well?

[philippnoah]

My bad. I must have looked at old output, I'm using spacy==3.1.0. Updated my previous comment.

[KTRosenberg]

My bad. I must have looked at old output, I'm using spacy==3.1.0. Updated my previous comment.

Still, would you show the steps? Also, you should be able to get it to 3.1.4

For 3.2 support, I think the author needs to do that explicitly.

[philippnoah]

Steps that work for me (just tested) are:

conda create -n coreferee python=3.9
conda activate coreferee
conda install -c apple tensorflow-deps==2.6.0
pip install tensorflow-macos==2.7.0
conda install -c conda-forge spacy==3.1.4
python -m spacy download fr_core_news_lg
pip install coreferee --no-deps
python -m coreferee install fr

Let me know.

[KTRosenberg]

Steps that work for me (just tested) are:

conda create -n coreferee python=3.9
conda activate coreferee
conda install -c apple tensorflow-deps==2.6.0
pip install tensorflow-macos==2.7.0
conda install -c conda-forge spacy==3.1.0
python -m spacy download fr_core_news_lg
pip install coreferee --no-deps
python -m coreferee install fr

Let me know.

Thanks. You may want to try 3.1.4 though. Also, make sure the transformers models work for you.

Does it matter if you use miniforge, mambaforge, etc. to install conda?

[philippnoah]

Just tested with spacy==3.1.4 and en_core_web_trf, worked for me.
I only tested with miniforge, I'm not sure if it works with the others.

[KTRosenberg]

For me I get a warning that there’a no CUDA device, which means it’s reverting to doinv everything on the CPU. I guess I have to reinstall everything somehow because that can’t be right.

[richardpaulhudson]

@KTRosenberg have you solved this now or are you still having problems? Working on the CPU is quite feasible as far as Coreferee is concerned because the neural network is much smaller than those used by typical transformer-based models.

[KTRosenberg]

@richardpaulhudson Hi. I got it working sort of, but only via miniforge/conda so I could use the newer tensorflow for Apple chips, and I have to install from sourcd
I still encounter issues with installation though, like numpy going out of sync. (I have to do an uninstall, then reinstall “hack.”) I am hoping coreferee might be upgraded to work with Spacy 3.2 because that woild also help.

My fear is that eventually 3.1.4 will stop being available, and then Ia will be stuck.

I do get much better results with the transformers version, so I am trying to stick with that one.

[richardpaulhudson]

Sorry about all this! I'm currently working on a new version of Coreferee which uses Thinc rather than TensorFlow, meaning that it doesn't have any dependencies except spaCy itself and the transitive dependencies that entails. I think this should solve these problems.

[KTRosenberg]

Sorry about all this! I'm currently working on a new version of Coreferee which uses Thinc rather than TensorFlow, meaning that it doesn't have any dependencies except spaCy itself and the transitive dependencies that entails. I think this should solve these problems.

Thanks for the update. I'm looking forward to that new version!

[richardpaulhudson]

Version 1.2.0 is now available and these problems should no longer be relevant.